Liuy write up Sha 2003 shallow parsing with conditional random fields

This is a review of sha_noyear_shallow_parsing_with_conditional_random_fields by user:Liuy.

This paper tries to train a CRF based on optimization algorithms and show better empirical results on shallow parsing. They show the log-linear sequence labeling discriminative model with optimization is effective for learning shallow parsers. It can do NP chunking better than other classifier combination by heuristics. However, I do not think they provide good reasons why and how this method can do equally well on some other tasks like POS. Thus it is not clear whether this approach can generalize its use in NLP problems.

Also I do not like their empirical study. It does not seem convincing to me. There are not clear enough explanation of the way they construct the dataset. The significance of their results and the intuition behind the seemingly advantage they achieved empirically are not given.